AppellateGen: leaderboard
Appellate legal judgment generation in Chinese (AppellateGen, Yang et al., 2026): given a first-instance judgment and the facts newly found on appeal, a single model drafts the second-instance judgment for 1,000 test cases, half reversed and half affirmed. A DeepSeek-V3.2 judge scores verdict consistency (zero when the ruling direction is wrong), fact consistency, legal application and logical reasoning on 0-5 rubrics, and the score is their mean; higher is better.
Metric: Mean LLM-Judge Score (0-5; verdict, fact, legal application, reasoning). Source: arxiv.org. Saturation forecast: Around July 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | QwQ-32B | 2.46 | #909 |
| 2 | Gemini 2.5 Flash (Non-reasoning) | 2.43 | #452 (Gemini 2.5 Flash) |
| 3 | Qwen 3 30B A3B 2507 (Thinking) | 2.3 | #759 |
| 4 | Qwen 3 30B A3B 2507 Instruct | 2.27 | #1078 |
| 5 | Qwen 3 235B A22B 2507 Instruct | 2.26 | #619 |
| 6 | Gemini 2.5 Flash Lite (Non-reasoning) | 2.1 | #958 (Gemini 2.5 Flash Lite) |
| 7 | DeepSeek R1 0528 Qwen3 8B | 2.09 | #2101 |
| 8 | Qwen 3 8B (Non-reasoning) | 1.82 | #1838 (Qwen 3 8B) |
| 9 | GPT-OSS-120B | 1.79 | #699 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=appellategen · How It Works · Data refreshed daily, snapshot 2026-11-12.